OpenAI's Control Room Can Halt AI Research Runs in 35 Seconds

OpenAI's Control Room halts model output across research infrastructure in 35 seconds, after a sandbox escape took 2.5 hours to stop manually. ChatGPT is not covered.